Professional Documents
Culture Documents
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/245188387
CITATIONS
READS
11
50
3 authors:
Carlos Alberto Cruz Villar
Jaime lvarez
SEE PROFILE
SEE PROFILE
Miguel G. Villarreal-Cervantes
Instituto Politcnico Nacional
25 PUBLICATIONS 66 CITATIONS
SEE PROFILE
Mechatronics
journal homepage: www.elsevier.com/locate/mechatronics
a r t i c l e
i n f o
Article history:
Received 15 May 2007
Accepted 1 September 2008
Keywords:
Concurrent redesign
Design optimization
Minimum time
Pendubot
a b s t r a c t
An interesting problem in robotics is to minimize the required time to force a manipulator to travel
between two specic points (positioning time). In this paper a concurrent structure-control redesign
approach is proposed in order to nd the minimum positioning time of an underactuated robot manipulator, by considering a synergetic combination between the structural parameters and a bangbang control law. The problem consists in nding the structural parameters of the system and the switching
intervals of the bangbang control that simultaneously minimize the positioning time, subject to the
input constraint and the structural parameter constraint. The concurrent structure-control redesign
approach is stated as a dynamic optimization problem (DOP). The projected gradient method is used
to solve the DOP.
The effectiveness of the proposed concurrent redesign approach is shown via simulation and experimental results on an underactuated system called the Pendubot.
2008 Elsevier Ltd. All rights reserved.
1. Introduction
Robots are currently used for a variety of tasks in industrial
applications. Improving their performance could result in commercial benets.
Time-optimal control of robots [1,2] has been used in order to
minimize the required time to take the system from an initial state
to a nal one (positioning time). To nd the time-optimal control,
innite dimensional problems are usually stated. A variational approach, which focuses on obtaining a solution via the classical necessary conditions for optimality, solves the innite dimensional
problem. It is proven in [4] that the behavior of the control signal
for time-optimal control problems with bounded control input and
with terminal constraints is of bangbang nature.1 Then, timeoptimal control problems can be parametrized [3] in order to
transform them into nite dimensional problems. So these nite
dimensional problems can be solved by using nonlinear programming (NLP) methods.
On the other hand, in the last years, an integrated design approach [58] which involves a structure-control design, has been
employed to improve the performance of electromechanical systems. This approach consists in integrating the design parameters
* Corresponding author. Tel.: +52 5550613800x6311.
E-mail addresses: gvillarr@cinvestav.mx, villamike@hotmail.com (M.G. Villarreal-Cervantes).
URL: http://www.meca.cinvestav.mx/ (M.G. Villarreal-Cervantes).
1
A control signal which is restricted to be in a lower and an upper bound and it
switches immediately from one extreme to another at certain times (the control
signal is never strictly between the bounds), is called bangbang control.
0957-4158/$ - see front matter 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.mechatronics.2008.09.002
179
J tf
x_ f xt; ut;
xt0 x0
xtf xf
umin 6 ut 6 umax
min 6 6 max
To transform the constrained problem stated above into an unconstrained one, barrier functions [10] are considered. These barrier
functions set a barrier against the solution leaves the feasible region
imposed by the constraints when they are added to the performance index of the optimization problem. Therefore, the dynamic
optimization problem stated in (1)(6) is reformulated in (8)(11),
which considers the inequality constraints (6) as barrier functions.
Moreover, dening the input control vector ut as a bangbang
control law, the switching time interval vector Dt of the control signal must be included in the next problem reformulation. Then ut
is stated in the following way:
ut
umax
if Dt j > 0
umin
if Dt j < 0
is
the
barrier
function
for
the
lmax
lmin
JDt; t f N
nDt
n
X
X
Dt j
Nl
j1
l1
x_ f xt; Dt;
xt 0 x0
10
ef xt f xf 0
11
T T 1
1 A
T
A
T A
if
n 6 nDt n ,
or
A
if
AA A AA
ef is the solution that minimizes the error
n > nDt n , and A
ef xtf xf . The convergence of (12) to a solution is guaranteed
with small step sizes kr and kn [10], that is, the constraint error ef
and the projection P DJ converge to zero. Then, the step sizes krii ; knii
are selected to be in the interval (0, 1]. In order to obtain the optimal parameters with less computing time, the constraint error fulllment ef must have more importance than the projection P DJ in
(12), that is, krii P knii must be taken into account
Dt new
Dt
ef kn PrJ
kr A
new
12
x_ j sgnDt j
o Dt j
ot j oDt j oxj
oxj
13
180
dAj t
of xt; Dt j ;
Aj t
ot
ox
14
d oxt
o
l
dt
of xt ; Dt; oxt
of xt; Dt;
oxt
o l
o l
15
q
1 ; q
2 T
the motor torque vector, q q1 ; q2 T ; q_ q_ 1 ; q_ 2 T and q
are the position, velocity and acceleration vectors of the links 1
and 2, respectively
_ q_ Gq
s Dqq Cq; q
16
Since in this work the nature of the open-loop control law is bang
bang and as the actuator (a DC armature controlled motor) can not
deliver a perfect bangbang signal, the dynamics of the motor (the
electrical constant time se and the mechanical constant time sm )
must be considered. Hence, the Pendubot dynamics (16) coupled
to the motor dynamics is taken into account and it can be found
in [14]. As it is stated in Section 2, a nonlinear state representation
of the system (2) is considered, where x q1 ; q2 ; q_ 1 ; q_ 2 , ia T is the
state vector, ia is the armature motor current and ut V in ; 0T is
the input voltage vector.
In Fig. 1, the schematic representation of the Pendubot is
shown, where m1 ; m2 ; l1 ; l2 ; lc1 , lc2 ; I1 ; I2 are the masses, lengths,
lengths between the rotation axis and the center of mass, and
the inertias of links 1 and 2, respectively.
To implement the concurrent redesign approach, it is necessary
to modify the structural parameters of the Pendubot. Then by considering that the Pendubot is already built, a weight of cylindrical
shape is placed at each link in order to modify the dynamic parameters mi ; lci ; Ii ; i 1; 2 of the Pendubot. These cylindrical weights
are shown in Fig. 2, where mp1 ; mp2 ; lcp1 ; lcp2 are the masses and
the distance between the rotation axis and the mass center of
cylindrical weights 1 and 2. Thus, mp1 Kg; mp2
Kg; lcp1 m; lcp2 mT 2 R4 is the structural parameter vector to
be modied.
The dynamic parameters of the Pendubot with the added cylindrical weights are obtained via classical methods [15], and summarized in Table 1. Note that these dynamic parameters
~ i; f
lci ; ~Ii ; i 1; 2 are functions of both the constant dynamic
m
parameters of the Pendubot m1 0:8435 Kg; m2 0:3859
Kg;lc1 0:1573 m;lc2 0:1416 m;I1 0:0057 Kgm2 ;I2 0:00539
Kgm2 and the parameters of the cylindrical weights which are
e 1 2 Kgm2 ;
grouped in , where Izp1 12 mp1 rp1 2 mp1 lcp1 lc
2
2
2
1
e
Kgm and rp1 0:0254 m,
Izp2 mp2 rp2 mp2 lcp2 lc 2
2
Fig. 2. CAD drawing of the Pendubot with the added cylindrical weights.
181
Link 2
~ i; f
the whole system m
lci ; ~Ii ; i 1; 2. Now, the motion equation of
the Pendubot x_ f xt; Dt; is a function of the state vector
xt, the switching time interval vector Dt of the control signal
ut and the structural parameter vector of the cylindrical
weights.
Changes in mpi for i 1; 2 must be positive because mass can
only be added to the already constructed Pendubot and changes
in lcpi for i 1; 2 could be positive or negative. A negative change
in lcpi means that the displacement of the cylindrical weight is
on the x-axis (see Fig. 2). Therefore, in order to establish the permissible changes in the structural parameter vector of the cylindrical weights, the inequality constraint min 6 6 max (6) is
introduced.
4. Simulation and experimental results
The swinging up problem of the redesigned Pendubot is solved
via the concurrent redesign approach described in Section 2. This
approach simultaneously nds both the switching interval vector
Dt Dt1 DtnDt 2 RnDt and the structural parameters of the
weights mp1 ; mp2 ; lcp1 ; lcp2 T 2 R4 that minimize JDt;
PnDt jDtj j P4 Nl subject to the motion equation of the
tf N
l1
j1
the
initial
position
Pendubot-Motor
x_ f xt; Dt; ,
T
xo p2 ; 0; 0; 0; 0 and the nal position xf p2 ; 0; 0; 0; 0T , assuming that there is a bangbang control ut dened in (7) where
umax 45 volts and umin 45 volts. The minimum and maximum
limit
vectors
min mp1min ; mp2min ; lcp1min ; lcp2min T ,
max mp1max ; mp2max ; lcp1max ; lcp2max T are explicitly dened as
and
max 0:2; 0:135; 0:127;
min 0; 0; 0:20; ;0:0381T
or
),
this
limit
value could
hits its limit value ( H
imin
imax
i
i
be enlarged if and only if this is physically possible. Then, other
optimization process which considers the modied limit vector
will be required for nding the optimum structural parameter vector. This procedure is repeated until the optimum structural
parameter vector H does not hit the limit vector or even when
it is not physically possible to enlarge the limit vector.
The constants of the barrier functions x and the step sizes
kr ; kn required in (12), are selected as in (17), where diag is
the diagonal matrix with the elements in the diagonal
x l 5E 8 for l 1 to 4
17
kr diag1E 3; . . . ; kr5;5 ; 3E 4; 3E 4; 5E 4; 5E 4
kn diag1E 4; . . . ; kn5;5 ; 1E 4; 1E 4; 1E 4; 1E 4
As the projected gradient algorithm could converge to a local minimum, several runs were performed in numerical simulations by
using different starting conditions ini ; Dt ini , which were randomly
generated inside the feasible region imposed by min and max . The
starting conditions that generated the lowest value of the performance index (8) result as indicated in (18) and (19). With those initial starting conditions, the optimal redesign structural parameters
and the optimal switching intervals are shown in (20) and (21)
18
19
T
20
T
21
With the optimum parameters (20) and (21), the Pendubot brings
the links from the rest position to the upwards position in a miniP5
H
mum nal time of t H
j1 jDt j j 1:09 s. In (20), it is important
f
to note that the cylindrical weight of the link 2 should not be
182
X1 (rad)
2
0
2
0
0.5
1.5
2.5
0.5
1.5
2.5
0.5
1.5
2.5
0.5
1.5
2.5
0.5
1.5
2.5
0.5
1.5
2.5
X2 (rad)
2
0
X3 (rad/s)
2
10
0
X4 (rad/s)
10
10
0
10
X5 (A)
5
0
5
50
u (V)
22
0
50
Fig. 4. Experimental results to swing up and stabilize the Pendubot with the concurrent redesign approach and a LQR control, respectively.
1:09 s
0:954 s
0:895 s
xt at the state
x xtf p2 ; 0; 0; 0; 0T . The LQR gains K pm are obtained by using a linearized model of the Pendubot around
the state
x and the lqrd Matlab function. Hence, K pm
403:5333, 421.7918, 91.8461, 64.0117, 2.9301] are the optimal LQR gains.
The switching of the control law from the one obtained by the C.
R. approach to the LQR, is carried out when the links are in the
attraction region of the LQR, given by, jex1 j 6 0:13963
rad; jex2 j 6 0:10472 rad; jex3 j 6 4 rad
and jex4 j 6 8 rad
, where
s
s
exi xi t xfi , which was experimentally obtained as the one that
minimizes the norm of the attraction region, given by the term
q
e2x1 e2x2 e2x3 e2x4 . Different attraction regions would lead to
different experimental positioning times. However, we are interested in the attraction region allowing the Pendubot to stay the
longer time being controlled by the bangbang control instead
the LQR control.
In Fig. 4, we show the experimental results of the C. R approach.
At the rst second, the open-loop bangbang control uDt with the
optimal switching interval vector (22) is applied. At time
t 0:895 s, the states of the Pendubot are inside the attraction region, then the LQR control replaces the bangbang control, stabilizing the links at
x. Then the required time to switch the control law
is 0:895 s. Due to the attraction region of the LQR, the nal time in
this experiment is shorter than t cH
f .
In Table 2 the summary of the simulation and experimental
results of the concurrent redesign approach is shown. The nal
time of experimental results is enhanced because the optimal
design parameters Dt cH and cH were modied when the gradient projection algorithm was computed again with the state
trajectory feedback from the real system, instead of the simulated trajectory. It should be pointed out that the experiment
was conducted several times. In order to give a statistical measure of the solution behavior, a fty sample set provided an
average nal time of t f 0:895 s and a standard deviation of
0:0204 s.
5. Conclusion
In this paper, a concurrent redesign approach that involves
structural and a bangbang control law parameters is proposed.
This approach reformulates the concurrent redesign problem as a
dynamic optimization problem, where the independent variables
are the structural (mechanical) and control parameters of an electromechanical system. To solve such a dynamic optimization problem, the projected gradient algorithm is used, therefore, all
functions involved in the redesign process, should be at least once
differentiable.
The concurrent redesign approach has been illustrated with a
highly nonlinear underactuated double pendulum system called
the Pendubot.
183